Model training methods, network fault prediction methods, devices, equipment, storage media, and computer program products

Abstract

The application provides a model training method, a network fault prediction method, a device, equipment, a storage medium and a computer program product. The model training method comprises: obtaining historical time series data and historical text data associated with a network; constructing a prompt set based on the historical time series data; selecting a target prompt from the prompt set, the target prompt being superior to the remaining prompts in the prompt set except the target prompt; training an initial model based on the target prompt, the historical time series data and the historical text data to obtain a target model, the target model being used for predicting network faults.

Description

Technical Field

[0001] This application relates to the field of wireless communication technology, and in particular to a model training method, a network fault prediction method, an apparatus, a device, a storage medium, and a computer program product. Background Technology

[0002] In the operation and maintenance of private communication networks, the network management platform is responsible for managing functions such as network configuration, faults, performance, security, and accounts. Among these, network fault detection is crucial. Network management platforms typically use threshold and alarm policies to push alarm logs to indicate network faults when indicators exceed thresholds and meet alarm policies, or they employ unsupervised machine learning algorithms for network anomaly detection.

[0003] However, among related technologies, fault detection techniques based on fixed thresholds cannot predict network failures in advance, while artificial intelligence-based solutions cannot adapt to randomly changing scenarios when predicting faults, resulting in low accuracy of network fault prediction results. Summary of the Invention

[0004] In view of this, the embodiments of this application aim to provide a model training method, a network fault prediction method, an apparatus, a device, a storage medium, and a computer program product that can improve the accuracy of network fault prediction results.

[0005] The technical solution of this application embodiment is implemented as follows:

[0006] In a first aspect, embodiments of this application provide a model training method, the method comprising:

[0007] Acquire historical time-series data and historical text data associated with the network;

[0008] Based on historical time-series data, a set of prompts is constructed; a target prompt is selected from the set of prompts, and the target prompt is preferred over the remaining prompts in the set other than the target prompt.

[0009] Based on the target prompts, historical time-series data, and historical text data, the initial model is trained to obtain the target model, which is used to predict network faults.

[0010] Secondly, embodiments of this application provide a network fault prediction method, the method comprising:

[0011] Obtain user operation information and network-related time-series data at the first moment;

[0012] The user operation information and time-series data are fused and processed to obtain the input prompt;

[0013] Network fault prediction is performed based on the input prompt and the target model, which is trained using the model training method described above.

[0014] Thirdly, embodiments of this application provide a model training apparatus, the model training apparatus comprising:

[0015] The first acquisition unit is used to acquire historical time-series data and historical text data associated with the network;

[0016] The selection unit is used to construct a set of prompts based on historical time-series data; and to select a target prompt from the set of prompts, with the target prompt being preferred over the remaining prompts in the set other than the target prompt.

[0017] The training unit is used to train the initial model based on the target prompt, historical time series data, and historical text data to obtain the target model, which is used to predict network faults.

[0018] Fourthly, embodiments of this application provide a network fault prediction device, the network fault prediction device comprising:

[0019] The second acquisition unit is used to acquire user operation information and time-series data associated with the network at the first moment.

[0020] The fusion unit is used to fuse user operation information and time-series data to obtain the input prompt;

[0021] The fault prediction unit is used to predict network faults based on the input prompt and the target model, which is trained using the model training method described above.

[0022] Fifthly, embodiments of this application provide a model training device, the model training device comprising: a first processor and a first memory; the first processor executes the running program stored in the first memory to implement the above-described model training method.

[0023] In a sixth aspect, embodiments of this application provide a network fault prediction device, the network fault prediction device comprising: a second processor and a second memory; the second processor implements the above-described network fault prediction method when executing a running program stored in the second memory.

[0024] In a seventh aspect, embodiments of this application provide a storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described model training method, or, when executed by a processor, implements the above-described network fault prediction method.

[0025] Eighthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the above-described model training method, or, when executed by a processor, implements the above-described network fault prediction method.

[0026] This application provides a model training method, a network fault prediction method, an apparatus, a device, a storage medium, and a computer program product. The model training method includes: acquiring historical time-series data and historical text data associated with the network; constructing a set of prompts based on the historical time-series data; selecting a target prompt from the set of prompts, wherein the target prompt is preferred over the remaining prompts in the set excluding the target prompt; training an initial model based on the target prompt, the historical time-series data, and the historical text data to obtain a target model, which is used to predict network faults. The network fault prediction method includes: acquiring user operation information and time-series data associated with the network at a first moment; fusing the user operation information and the time-series data to obtain a prompt to be input; and performing network fault prediction based on the prompt to be input and the target model, wherein the target model is trained using the above model training method. By adopting the above implementation scheme, after constructing a set of prompts using historical time-series data, the prompt with the best performance is selected as the target prompt. When training the initial model, the target prompt is used in combination with historical time-series data and historical text data. Since the target prompt has the best performance, the predictive ability of the target model can be improved. When using the target model trained based on historical time-series data for network fault prediction, the prediction results obtained by inputting user operation information and time-series data at different times can match the actual network situation at different times, thus making the accuracy of network fault prediction higher. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of a model training method provided in an embodiment of this application;

[0028] Figure 2 This is an exemplary schematic diagram illustrating reasoning using a Large Language Model (LLM) as provided in an embodiment of this application.

[0029] Figure 3 This is a schematic diagram of the overall process of a model training method provided in an embodiment of this application;

[0030] Figure 4 This is a schematic flowchart of a network fault prediction method provided in an embodiment of this application;

[0031] Figure 5A schematic diagram illustrating an exemplary method for forming a prompt prefix, provided for embodiments of this application;

[0032] Figure 6 An exemplary input-output interaction flowchart of a target model is provided for embodiments of this application;

[0033] Figure 7 This is a schematic diagram of the overall process of a network fault prediction method provided in an embodiment of this application;

[0034] Figure 8 A schematic diagram illustrating the overall process of a model training method and a network fault prediction method provided in this application embodiment;

[0035] Figure 9 This is an illustrative diagram showing the comparison results of predicting the first screen latency;

[0036] Figure 10 This is a schematic diagram of the composition structure of a model training device provided in an embodiment of this application;

[0037] Figure 11 This is a schematic diagram of the composition structure of a model training device provided in an embodiment of this application;

[0038] Figure 12 This is a schematic diagram of the composition structure of a network fault prediction device provided in an embodiment of this application;

[0039] Figure 13 This is a schematic diagram of the composition structure of a network fault prediction device provided in an embodiment of this application. Detailed Implementation

[0040] To gain a more detailed understanding of the features and technical content of the embodiments of this application, the technical solution of this application will be further described in detail below with reference to the accompanying drawings and specific embodiments. The accompanying drawings are for reference only and are not intended to limit the embodiments of this application.

[0041] Unless otherwise defined, all technical and scientific terms used in the embodiments of this application have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the embodiments of this application is for the purpose of describing the embodiments of this application only and is not intended to limit this application.

[0042] In the following description, references to "some embodiments" refer to a subset of all possible embodiments. It is understood that "some embodiments" may be the same or different subsets of all possible embodiments and may be combined with each other without conflict. It should also be noted that the terms "first / second / third" used in the embodiments of this application are merely for distinguishing similar objects and do not represent a specific ordering of objects. It is understood that "first / second / third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein.

[0043] In the operation and maintenance of private communication networks, the network management platform (also known as the network management system) is responsible for managing functions such as network configuration, network faults, network performance, network security, and network accounts. Among these, network fault detection is crucial in the operation and maintenance of private communication networks. The network management platform typically sets thresholds and alarm policies. When an indicator exceeds the threshold and meets the alarm policy, an alarm log is pushed to indicate that a network fault has occurred. In addition to setting fixed thresholds for alarms, related technologies also employ unsupervised machine learning algorithms for anomaly detection, providing threshold-free fault identification capabilities. Simultaneously, trend predictions for key operation and maintenance indicators are generated using triple exponential smoothing or regression algorithms. These predictions are combined with business call chains, anomaly causal relationships, or expert rules to perform predictive analysis and early warning of faults. Alternatively, a system log fault feature library can be built, and system health profiles can be created based on classification algorithms. Multimodal learning can be used for power system fault early warning.

[0044] Based on the above technical solutions, fault detection technology based on fixed thresholds cannot predict the occurrence of faults in advance. When using artificial intelligence (AI) technology for fault prediction, it is generally necessary to train the model with historical experience data. The model cannot continuously update user operation information. That is, in the network fault prediction stage, changes in non-prediction-related indicators are not considered, such as changes in the number of terminals (e.g., data isolation in private network campuses makes it difficult to obtain corresponding training data), changes in terminal services (e.g., browsing and downloading services / video services), changes in data center temperature, and other external information.

[0045] In related technologies, industry network management platforms typically rely on contacting maintenance engineers after a fault occurs to predict its occurrence, resulting in lengthy resolution cycles. Therefore, achieving intelligent and automated fault prediction for private communication networks is of great significance.

[0046] To address the aforementioned problems, embodiments of this application provide a model training method, such as... Figure 1 As shown, the method may include:

[0047] S101. Obtain historical time-series data and historical text data associated with the network.

[0048] In this embodiment of the application, the model training method belongs to the data fine-tuning stage. Historical time series data and historical text data are the data sources used in the data fine-tuning stage. Historical time series data are stored in the historical time series data set (or historical time series data table).

[0049] In this embodiment of the application, historical text data includes text data in the historical communication private network knowledge base and text data in the historical communication private network operation and maintenance experience base.

[0050] In this embodiment of the application, the text data in the historical communication private network knowledge base is collected communication industry expertise in text format.

[0051] In this embodiment of the application, the text data in the historical communication private network operation and maintenance experience database includes communication private network operation and maintenance experience in text format provided by operation and maintenance experts and alarm record tables in data table format.

[0052] In this embodiment of the application, the historical time series data stored in the historical time series data table used in the data fine-tuning stage includes the field data to be predicted and the field data related to the field to be predicted. For example, if the field data to be predicted is the first screen latency, then the field data related to the first screen latency includes indicators such as the number of terminals, TCP connection establishment latency, and the number of HTTP no-response times, as shown in Table 1 below. Table 1 exemplarily represents a historical time series data table, in which the fields include time, business scenario, number of sessions, Transmission Control Protocol (TCP) connection establishment latency, client round-trip time (RTT) latency, number of Hypertext Transfer Protocol (HTTP) no-response times, first screen latency, etc.

[0053] Table 1

[0054]

[0055] S102. Based on historical time series data, construct a set of prompts; select the target prompt from the set of prompts.

[0056] Among them, the target prompt is superior to the remaining prompts in the prompt set other than the target prompt.

[0057] In this embodiment of the application, a processing module A is designed to implement the process of constructing prompts and selecting target prompts.

[0058] It should be noted that the naming of processing module A can be selected according to the actual situation, and no specific limitation is made in this embodiment.

[0059] In this embodiment, the prompts are initialized by the prompt initialization module of processing module A. The corresponding prompts are obtained from the historical time-series data in the historical time-series data table through manual writing. Furthermore, the initially written prompts can be further enriched through manual writing, forming a prompt set from all the written prompts. Table 2 below lists some examples of the prompts obtained after initialization.

[0060] Table 2

[0061]

[0062] In this embodiment of the application, in addition to enriching the types of prompts by manually writing them, the word order of the prompts can be adjusted by LLM (Limited Language Management) to obtain even more types of prompts.

[0063] It should be noted that, in addition to using LLM as described above, other adjustment methods may also be adopted, and no specific limitations are made in the embodiments of this application.

[0064] In this embodiment of the application, after obtaining the set of prompts, the processing module A further selects the optimal prompt from the set of prompts.

[0065] In this embodiment of the application, selecting the optimal prompt from the prompt set can be done by selecting the one with the highest score from the prompt set based on the score of each prompt in the prompt set.

[0066] It should be noted that the target prompt can be understood as a template input into the target model during training. For example, when predicting the first screen latency, the historical first screen latency data can be used to replace the first screen latency-related data in the target prompt.

[0067] S103. Based on the target prompt, historical time series data and historical text data, train the initial model to obtain the target model.

[0068] The target model is used to predict network faults.

[0069] In this embodiment of the application, the target model is obtained by training the initial model based on the target prompt, historical time series data and historical text data. This can be achieved by: constructing question-answer pairs based on historical text data; and training the initial model based on the target prompt, historical time series data and question-answer pairs to obtain the target model.

[0070] In the embodiments of this application, the initial model can be an initial LLM model.

[0071] In this embodiment of the application, the construction of question-answer pairs based on historical text data can be achieved through the designed processing module B.

[0072] It should be noted that the naming of processing module B can be selected according to the actual situation, and no specific limitation is made in this embodiment.

[0073] In this embodiment of the application, when processing module B processes historical text data and obtains question-and-answer pairs, it can form question-and-answer pairs based on the text-formatted communication industry knowledge base in the input historical communication private network knowledge base and / or the text-formatted communication private network operation and maintenance experience feedback from operation and maintenance experts in the historical communication private network operation and maintenance experience base and the alarm record table in the form of a data table.

[0074] In this embodiment of the application, when training the initial model, the historical time series data corresponding to the historical time series data table is input into the target prompt, and the question-and-answer pair formed by using historical text data is simultaneously input into the initial LLM model for training, so as to obtain the target model for network fault prediction.

[0075] It is understood that, in the model training method provided in this application embodiment, after constructing a set of prompts using historical time-series data, the method selects the prompt with the best performance from the set as the target prompt. When training the initial model, the target prompt is used in combination with historical time-series data and historical text data for training. Because the target prompt has the best performance, the predictive ability of the trained target model can be improved. When using the target model for prediction, the accuracy of the prediction results can be improved. This ensures that when using the trained target model to predict network faults based on user operation information and time-series data at different times, the prediction results obtained match the actual network situation at different times as the user operation information and time-series data change, resulting in a higher accuracy of network fault prediction.

[0076] In one embodiment of this application, a target prompt is selected from a set of prompts. Specifically, each prompt in the set of prompts can be scored; multiple prompts can be selected from the set of prompts, with multiple prompts selected each time; for each of the multiple selected prompts, the prompt with the highest score is selected from the multiple prompts selected each time, resulting in multiple prompts with the highest scores; and the target prompt is selected from the multiple prompts with the highest scores.

[0077] In this embodiment, each prompt contains historical time-series data at n time points, where n is an integer greater than or equal to 1. When scoring each prompt in the prompt set, based on the historical time-series data at the n time points contained in each prompt, the time-series data at the (n+1)th time point is predicted to obtain predicted data; the predicted data and the historical time-series data at the (n+1)th time point are processed to obtain a first value; and a second value is determined based on the historical time-series data at the n time points and the historical time-series data at the (n+1)th time point; based on the first and second values, the score corresponding to each prompt in the prompt set is determined.

[0078] In this embodiment of the application, the prompt set contains multiple prompts, and each prompt contains historical time series data at n time points.

[0079] In the embodiments of this application, the first value can represent the error value E, and the second value can represent the absolute difference H.

[0080] In this embodiment of the application, scoring each prompt in the prompt set can be achieved through the evaluation module included in the processing module A.

[0081] In this embodiment of the application, the evaluation module evaluates either a set of manually written prompts or a set of prompts P obtained by further rewriting and adjusting the manually written prompts. 1...m For each prompt P i The module was evaluated for effectiveness and ultimately received a score of P. 1...m Output in the form of {[Prompt 1, Score]...[Prompt m, Score]}.

[0082] In this embodiment of the application, the evaluation module first performs inference through LLM, that is, the process of obtaining prediction data.

[0083] In this embodiment, since each prompt contains historical time series data at n time points, the prompts included in the prediction instruction I are input into the LLM. That is, the indicator data (i.e., historical time series data) X at time points 1...n in the prompts included in the prediction instruction I are used to predict the indicator data (i.e., prediction data) corresponding to time point n+1, and the prediction data X corresponding to time point n+1 is obtained. n ′ +1 .

[0084] In this embodiment of the application, the reasoning process for predicting data can be performed from a single dimension or multiple dimensions. For example, a single dimension can only predict the data at time point n+1 corresponding to the first screen delay, while a multiple dimension can predict the data at time n+1 corresponding to the first screen delay, and can also predict the data at time n+1 corresponding to TCP connection establishment delay, etc. The following formula (1) shows the process of prediction reasoning in a single dimension, and formula (2) shows the process of prediction reasoning in a multiple dimension.

[0085] X = [x1, x2, x3, x4, ..., x n ,x n+1 ],I(X'=[x1,x2,x3,x4,...,x n ])→LLM→x' n+1 (1)

[0086] X=[x1,x2,x3,x4,...,x n x n+1 ], Y=[y1, y2, y3, y4,..., y n y n+1 ],...,Z=[z1,z2,z3,z4,...,z n , z n+1 ]

[0087] I(X1' ...n ,Y1' ...n Z' 1...n )→LLM→x' n+1 (2)

[0089] In formula (1), the X set represents the historical time series data corresponding to the n+1 time points contained in the historical time series data, I represents the prediction instruction, and the X′ set represents the historical time series data corresponding to the first n time points, x′ n+1 This refers to prediction data obtained from a single dimension, such as the predicted first-screen latency.

[0090] In formula (2), set X represents the historical time series data corresponding to the n+1 time points in the first dimension, set Y represents the historical time series data corresponding to the n+1 time points in the second dimension, set Z represents the historical time series data corresponding to the n+1 time points in the third dimension, I represents the prediction instruction, set X′ represents the historical time series data corresponding to the first n time points in the first dimension, set Y′ represents the historical time series data corresponding to the first n time points in the second dimension, set Z′ represents the historical time series data corresponding to the first n time points in the third dimension, and x′ represents the historical time series data corresponding to the first n time points in the third dimension. n+1 This refers to prediction data obtained from multiple dimensions, such as predicted first-screen latency, TCP connection establishment latency, and server-side RTT latency.

[0091] In this embodiment of the application, after obtaining the predicted data corresponding to the (n+1)th time point, the error value E and the absolute error H are calculated by combining the true value of the historical time series data corresponding to the (n+1)th time point.

[0092] In this embodiment of the application, the predicted data x' at the (n+1)th time point is calculated. n+1 and the true value (i.e., the historical time series data at the (n+1)th time point) x n+1 The error value E between them can be calculated using the following formula (3).

[0093] E = |x' n+1 -x n+1 | (3)

[0094] In this embodiment of the application, for each prompt, the absolute difference H is determined by the average value of the indicator data (i.e., historical time series data) at time points 1...n in the prompt and the historical time series data at the (n+1)th time point, which can be calculated using the following formula (4). Formula (4) can be used to calculate the average value of the indicator data (i.e., historical time series data) at time points 1...n and x. n+1 The absolute difference H.

[0095]

[0096] It should be noted that since the data used here is historical time series data, the historical time series data at the (n+1)th time point can be obtained directly.

[0097] In this embodiment, when determining the score for each prompt based on the error value E and the absolute difference H, the score can be calculated as shown in the following formula (5). Wherein, if E is less than H, the score is 10 + 1 / MAE(x' n+1 x n+1If E equals H, the score is 5; if E is greater than H, the score is 0.

[0098]

[0099] Where MAE represents the mean absolute error, which can be expressed in English as Mean Absolute Error, 1 / MAE(x') n+1 x n+1 ) is equivalent to

[0100] In this embodiment of the application, after obtaining the score corresponding to each prompt in the prompt set, the target prompt is further selected by the selection module included in the processing module A. Specifically, for the prompt set P... 1...m The corresponding error score Rank 1...m For the set of prompts P 1...m Perform h selections, each time from the prompt set P 1...m S randomly sample s prompts (with replacement), and then select the prompt with the best score from the s prompts and add it to the list PARENTS. 1...h In this selection process, a prompt is chosen only if its error score is better than the other s-1 competitors out of s prompts. If the error scores are the same, the longer prompt is chosen because it contains more information and is therefore considered superior.

[0101] In this embodiment, since the highest-rated option is selected from s individual prompts each time, and h selections are performed, the final output of the selection module is the list PARENTS. 1...h {[Prompt 1, rating]...[Prompt h, rating]} represents the h highest-rated prompts obtained from h selections.

[0102] In this embodiment of the application, the target prompt determined from the h prompts is the prompt with the highest score among the h prompts.

[0103] For example, if h is 3 and s is 5, and the highest scores of the prompts obtained from the three selections are 11.5, 10.8, and 11.9 respectively, then the target prompt is the one with the highest score among these three, that is, the prompt corresponding to 11.9 is taken as the target prompt. And 11.5 is the prompt with the highest score among the five prompts selected each time, and so on, which will not be elaborated here.

[0104] In one embodiment of this application, for each selected prompt, the prompt with the highest rating is chosen from the multiple selected prompts. After obtaining multiple prompts with the highest ratings, these prompts can be recombine to obtain M combined prompts, where M is an integer greater than or equal to 0. The M combined prompts are then subjected to word order mutation processing to obtain M word order adjusted prompts. Based on the multiple prompts with the highest ratings and the M word order adjusted prompts, an updated prompt set is constructed. The updated prompt set is used to select the target prompt from the updated prompt set.

[0105] In this embodiment of the application, for each selection of multiple prompts, the prompt with the highest score is selected from the multiple prompts selected each time. After obtaining multiple prompts with the highest scores, the evolution module included in the processing module A can be used to further recombine and perform word order mutation processing on the multiple prompts with the highest scores.

[0106] In this embodiment of the application, the selected h prompts, namely PARENTS 1...h The prompts in the code can be further expanded using two evolutionary operators to guide the expansion. The final output of this evolutionary module is PARENTS. 1...h OFFSPRING generated after word order mutation 1...m-h .

[0107] It should be noted that since the initial set of prompts contains m prompts, after the word order mutation, the output PARENTS will be... 1...h OFFSPRING generated after word order mutation 1...m-h The sum of these is also m.

[0108] In this embodiment of the application, the h prompts are first recombined. The recombination can be achieved by cross-transformation, that is, two prompts are randomly selected from the h prompts as context, and the LLM is input to rephrase and generate two new prompts. This process is repeated until a total of M = mh new prompts are generated.

[0109] It should be noted that LLM is just one example model, and other models can also be used.

[0110] In the embodiments of this application, the mh new prompts obtained after cross-transformation, together with the h prompts, constitute m prompts.

[0111] In this embodiment, mh prompts are subjected to word order mutation processing to obtain mh prompts with adjusted word order. Specifically, for each new prompt after cross-transformation, a mutation method is randomly selected to adjust the word order or content.

[0112] It should be noted that the word order mutation processing only adjusts the word order of the prompts and does not increase the number of prompts.

[0113] In this application embodiment, four methods for enriching the word order are designed, which include nine specific methods to provide mutation guidance for the diversification of prompts. The first method is multivariate numerical description, which makes the learning of the large model LLM independent of specific numbers by changing the numerical description of the question. Specific methods include Chinese expression replacement, integer and decimal replacement, and changing the position of the decimal point. The second method is numerical instruction adjustment, which includes time series progression and prediction cycle changes. The third method is prediction field adjustment, which includes related field adjustment and prediction target adjustment. The fourth method is context adjustment, which includes Few-shot addition and CoT addition. The four methods for enriching the word order and the nine specific methods are shown in Table 3 below.

[0114] Table 3

[0115]

[0116]

[0117] In this embodiment, the mh prompts output after the evolution module is recombined and the word order mutation is processed, along with the h highest-rated prompts, are merged into a new prompt set P. 1...m The evaluation module is then re-entered for iterative evaluation. After reaching the maximum number of iterations, the selection module continues to select, and finally selects the prompt with the highest error score from the selected h prompts as the excellent recommendation text (i.e., the target prompt). Using this target prompt as a template, the historical time series data in the historical time series data table is organized and further input into the LLM for retraining.

[0118] It should be noted that the maximum number of iterations can be selected according to the actual situation, and no specific limitation is made in this embodiment.

[0119] In this embodiment of the application, during the training of LLM, the input of LLM includes: organizing historical time-series data using target prompts and using the organized prompts as LLM inputs. In addition, it also includes: using question-answer pairs obtained by processing text data from the historical communication private network knowledge base and text data from the historical communication private network operation and maintenance experience base as LLM inputs.

[0120] In this embodiment, the target prompts, processed from historical time-series data, and question-answer pairs are used to perform low-rank adaptive fine-tuning of large language models (LORA). A bypass is added alongside the initial LLM to perform dimensionality reduction and then dimensionality increase. During training, the parameters of the pre-trained initial LLM are fixed, and only the dimensionality reduction and dimensionality increase matrices are trained. The input and output dimensions of the initial LLM remain unchanged. At output, the two matrices are superimposed with the parameters of the pre-trained initial model. The final target model is obtained through parameter updates and adjustments. The target model can be understood as a fault prediction model, which predicts network faults.

[0121] Based on the above embodiments, since LLM responds intelligently based on prompt input, different prompts can significantly affect LLM's reasoning ability, and the LLM reasoning result is as follows: Figure 2 As shown, Figure 2 The input command is a simple subtask in the prediction task: sequence segmentation. Figure 2 The four different prompts for the task presented correspond to four different answers given by the same LLM, demonstrating that the LLM responded differently to different descriptions of the same task. It can be seen that a high-quality fine-tuning dataset (i.e., selecting multiple prompts with the highest scores) can improve the prediction accuracy of the LLM during the inference phase. Therefore, to select diverse, high-quality, and abundant datasets during the data fine-tuning phase, processing module A was designed in this embodiment. This processing module A redesigned the prompt selection process based on the principles of evolutionary algorithms. The overall flow of processing module A is as follows: Figure 3 As shown, basic LLMs inherently possess very weak predictive capabilities, such as relying on simple calculations like averages to estimate data. However, due to the very weak mathematical capabilities of basic LLMs, even simple averaging predictions frequently result in calculation errors. To improve the predictive capabilities of basic LLMs, this embodiment incorporates a multi-step selection process using different processing modules during the fine-tuning stage. This optimizes the input prompt templates to the LLM, thereby updating the network parameters. Processing module A primarily includes a prompt initialization module, an evaluation module, a selection module, and an evolution module. Ultimately, the selection module obtains the target prompt (i.e., the recommended text), and historical time-series data is organized into training data based on the target prompt and input into the LLM for training. This can be understood as fine-tuning the model parameters of the LLM model.

[0122] It should be noted that each processing step of the above processing module A has been described in detail in the foregoing embodiments, and will not be repeated here.

[0123] This application also provides a network fault prediction method, such as... Figure 4 As shown, the method may include:

[0124] S401. Obtain user operation information and time sequence data associated with the network at the first moment.

[0125] In this embodiment of the application, the first moment is the moment when the user performs information operation. The first moment can also be understood as the current moment, which is a real-time acquisition process.

[0126] In the embodiments of this application, network fault prediction can be understood as the reasoning stage.

[0127] In this embodiment of the application, the user operation information includes unstructured text information extracted from the user operation analysis module, which does not correspond to the time series data table.

[0128] In the embodiments of this application, the time-series data associated with the network at the first moment can be understood as the time-series data in the time-series data table that is updated in real time.

[0129] S402. The user operation information and timing data are fused and processed to obtain the input prompt.

[0130] In this embodiment of the application, the fusion processing of user operation information and time-series data to obtain the input prompt can be achieved in the following way: the user operation information is transformed to form a first input prompt; a prompt template is obtained, and based on the prompt template, a second input prompt corresponding to the time-series data is determined; the first input prompt and the second input prompt are fused to obtain the input prompt.

[0131] In this embodiment of the application, the first input prompt can be understood as a prompt prefix.

[0132] In this embodiment of the application, after selecting the target prompt, the target prompt can be stored in a fixed location. When network fault prediction is required, the target prompt is obtained from the location where the target prompt is stored, and the time series data to be predicted in the real-time updated data table is organized using the target prompt to obtain the second prompt to be input corresponding to the time series data. This can be understood as organizing the time series data into the form of the target prompt, and inputting the second prompt to be input into the processing module C.

[0133] In this embodiment, when a user adds a new terminal or modifies the network protocol to adjust terminal services, the network management platform requires the user to manually configure relevant information. The user operation analysis module obtains the user's current intent by having the user actively add information from the front end, and performs text conversion based on the user's current intent to form a prompt prefix.

[0134] In this embodiment of the application, the prompt prefix can be added to the processed prompt context during the reasoning stage, such as... Figure 5 As shown, the fused input prompt can be used as input to the target model. Free text input is an optional operation. Users can freely add prefixes to affect the fault prediction results, so that the target model can take into account real-time changes in the external environment during the inference stage. For example, if the computer room air conditioning fails, the range in which the computer room temperature may rise can be manually entered.

[0135] S403. Predict network faults based on input prompts and target models.

[0136] The target model was trained using the model training method described above.

[0137] In this embodiment, the processing module C fuses the second input prompt corresponding to the time series data in the real-time acquired time series data table and the prompt prefix corresponding to the user operation information in the user operation analysis module to obtain the fused input prompt. The input prompt is used as the input to the target model. The target model performs inference based on the text content in the input prompt to predict network faults.

[0138] In this embodiment of the application, when predicting network faults, the prediction accuracy is higher by taking into account the dimension of user operation.

[0139] For example, such as Figure 6 The diagram shows some of the input-output interaction flowcharts for the target model (also known as the fault prediction model).

[0140] It is understood that, in the network fault prediction method provided in this application embodiment, when using a target model trained based on historical time series data to predict network faults, the prediction results obtained by inputting user operation information and time series data at different times can match the actual network situation at different times, thus making the accuracy of network fault prediction higher.

[0141] Based on this, due to the complexity of network operation and maintenance in the 2B (Business) scenario of 5G, it is necessary to build an intelligent operation and maintenance capability system. The embodiments of this application can be used as fault prediction capabilities for the network operation and maintenance management platform of the communication private network, and can also be applied to other scenarios that require fault prediction capabilities.

[0142] In one embodiment of this application, after predicting network faults based on the input prompt and the target model, the prediction results obtained by using the target model, user operation information, and alarm information are fused and output.

[0143] In this embodiment of the application, the network fault inference stage is carried out using the target model. Finally, the prediction results, user operation information and potential alarm information are fused together through a preset template and output to the network management platform in the form of a fault prediction report.

[0144] It should be noted that the preset template can be selected according to the actual situation, and no specific limitation is made in this embodiment.

[0145] Based on the above embodiments, the overall flowchart of the inference stage in this application embodiment is as follows: Figure 7 As shown, continuously updated user operation information is added during the fault prediction process of the private network communication network. The specific processing procedures involved can be referred to the implementation process in the above embodiments, and will not be repeated here.

[0146] Based on the above embodiments, this application also provides an overall flowchart of a model training method and a network fault prediction method, as shown below. Figure 8 As shown, by leveraging the sequence modeling and text processing capabilities of LLM, and by learning from historical time-series data, text data from historical communication network knowledge bases, and text data from historical communication network operation and maintenance experience bases, and combining LLM's contextual learning capabilities with user operation information, network faults in communication networks can be predicted.

[0147] Specifically, this application embodiment designs a system in which multiple modules, including an offline phase (i.e., a data fine-tuning phase) and an inference phase, work together to predict faults in the private communication network. The overall implementation process is as follows: Figure 8As shown. Data fine-tuning involves training some or all of the parameters of the initial LLM. In the offline phase, input historical time-series data related to prediction, text data from the industry network knowledge base, and operation and maintenance experience base are processed by processing modules A and B to form two types of prompts to fine-tune the initial LLM. The text data uses manually labeled data and supervised fine-tuning (SFT) to improve the basic large model's ability in fault prediction using a question (Q)-answer (A) combination. Processing module A integrates historical time-series data and combines evolutionary algorithms to optimize the search for the target prompt. The initial LLM can use open-source tools such as ChatGLM2-6B and Qwen-7B. In the inference phase, based on real-time updated prediction-related time-series data and user operation information collected by the user operation analysis module, processing module C integrates continuously updated prediction time-series data and unstructured user operation information to form the LLM (i.e., the target model) after fine-tuning the input prompt. The target model outputs a fault prediction report or provides feedback to the network management platform in the form of an alarm. The specific implementation process involved can be referred to in the aforementioned embodiments, and will not be repeated here.

[0148] Based on the foregoing embodiments, this application also provides a comparison between an exemplary diagram of network fault prediction in related technologies and a diagram of network fault prediction combined with real-time updated user operation information in this application embodiment, such as... Figure 9 As shown. In Figure 9 As can be seen from the data, Day 1 shows the historical first-screen latency obtained using related technologies, while Day 2 includes the results of first-screen latency prediction using related technologies and the results of first-screen latency prediction by adding unstructured user operation information (or environmental information) in the embodiments of this application.

[0149] In this embodiment, because the large model (i.e., the LLM model) has stronger generalization ability than other AI models, it can be used across domains during the inference phase. For example, during model training, the training data prediction may not consider temperature data, such as the impact of data center temperature on predicted network latency. However, during the inference phase, the user inputs temperature information into the network management platform, and the LLM model can make a corresponding judgment. Other AI models require temperature information in their training data so that they can draw conclusions based on temperature information during the inference phase. Therefore, this embodiment can be applied to changing scenarios.

[0150] Compared with related technologies, the technical solution of this application has the following technical advantages:

[0151] 1. Compared with related technical solutions that use unsupervised algorithms combined with regression algorithms for fault detection and prediction, the embodiments of this application add external user operation information through processing module C to assist the target model in predicting fault indicators during the inference stage of fault prediction, resulting in more accurate prediction results.

[0152] 2. Compared with the fault prediction methods of related technical solutions, the reasoning process of the embodiments of this application is more flexible, and unstructured text information can be added as the context content of the target model during the reasoning stage.

[0153] 3. The embodiments of this application design a multi-step evolutionary search method (i.e., processing module A) to improve the training efficiency of supervised fine-tuning of large models.

[0154] Based on the above embodiments, another embodiment of this application provides a model training device 1, such as... Figure 10 As shown, the model training device 1 includes:

[0155] The first acquisition unit 10 is used to acquire historical time-series data and historical text data associated with the network.

[0156] Selection unit 11 is used to construct a set of prompts based on historical time series data; and to select a target prompt from the set of prompts, wherein the target prompt is superior to the remaining prompts in the set other than the target prompt.

[0157] Training unit 12 is used to train the initial model based on the target prompt, historical time series data and historical text data to obtain the target model, which is used to predict network faults.

[0158] In one embodiment, the model training device 1 may further include a scoring unit.

[0159] The scoring unit is used to score each prompt in the set of prompts.

[0160] Selection unit 11 is also used to select multiple prompts from the prompt set, selecting multiple prompts from the prompt set each time; for each selected prompt, selecting the prompt with the highest score from the selected prompts to obtain multiple prompts with the highest scores; and selecting the target prompt from the multiple prompts with the highest scores.

[0161] In one embodiment, each prompt contains historical time-series data for n time points, where n is an integer greater than or equal to 1.

[0162] In one embodiment, the model training device 1 may further include: a prediction unit and a processing unit.

[0163] The prediction unit is used to predict the time series data of the (n+1)th time point based on the historical time series data of n time points contained in each prompt, and obtain the predicted data.

[0164] The processing unit is used to perform calculations on the predicted data and the historical time series data at the (n+1)th time point to obtain a first value; and to determine a second value based on the historical time series data at the nth time point and the historical time series data at the (n+1)th time point.

[0165] The scoring unit is also used to determine the score corresponding to each prompt in the prompt set based on the first value and the second value.

[0166] In one embodiment, the model training device 1 may further include: a combination unit and a construction unit.

[0167] The combination unit is used to recombine multiple highest-rated prompts to obtain M combined prompts, where M is an integer greater than or equal to 0.

[0168] The processing unit is also used to perform word order mutation processing on the M combined prompts to obtain M prompts with adjusted word order.

[0169] The construction unit is used to build an updated set of prompts based on multiple highest-rated prompts and M prompts with adjusted word order. The updated set of prompts is used to select the target prompt from the updated set of prompts.

[0170] In one embodiment, the construction unit is also used to construct question-answer pairs based on historical text data.

[0171] Training unit 12 is also used to train the initial model based on the target prompt, historical time series data and question-answer pairs to obtain the target model.

[0172] This application provides a model training device that acquires historical time-series data and historical text data associated with a network; constructs a set of prompts based on the historical time-series data; selects a target prompt from the prompt set, where the target prompt is superior to the remaining prompts in the prompt set; and trains an initial model based on the target prompt, historical time-series data, and historical text data to obtain a target model. The target model is used to predict network faults. Therefore, this application's model training device, after constructing a set of prompts using historical time-series data, selects the best-performing prompt as the target prompt. When training the initial model, the target prompt is combined with historical time-series data and historical text data. Because the target prompt has the best performance, the predictive ability of the trained target model is improved. When using the target model for prediction, the accuracy of the prediction results is improved. This ensures that when using the trained target model to predict network faults based on user operation information and time-series data at different times, the prediction results match the actual network situation at different times, resulting in higher accuracy in network fault prediction.

[0173] Figure 11 This is a schematic diagram of the composition structure of a model training device 2 provided in an embodiment of this application. In practical applications, based on the same disclosed concept of the above embodiments, such as... Figure 11 As shown, the model training device 2 in this embodiment includes a first processor 20, a first memory 21, and a first communication bus 22.

[0174] In specific embodiments, the first acquisition unit 10, selection unit 11, training unit 12, scoring unit, prediction unit, processing unit, combination unit, and construction unit described above can be implemented by a first processor 20 located on the model training device 2. The first processor 20 can be at least one of the following: Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Digital Signal Processing Device (DSPD), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), CPU, controller, microcontroller, and microprocessor. It is understood that for different devices, the electronic devices used to implement the above processor functions can also be other types, and this application embodiment does not impose specific limitations.

[0175] In this embodiment, the first communication bus 22 is used to establish a connection between the first processor 20 and the first memory 21; when the first processor 20 executes the running program stored in the first memory 21, it implements the following model training method:

[0176] Acquire historical time-series data and historical text data associated with the network; construct a set of prompts based on the historical time-series data; select a target prompt from the prompt set, with the target prompt being preferred over the remaining prompts in the prompt set; train the initial model based on the target prompt, historical time-series data, and historical text data to obtain the target model, which is used to predict network faults.

[0177] In one embodiment, the first processor 20 is further configured to score each prompt in the prompt set; select multiple prompts from the prompt set, selecting multiple prompts from the prompt set each time; for each selected prompt, select the prompt with the highest score from the selected prompts to obtain multiple prompts with the highest scores; and select a target prompt from the multiple prompts with the highest scores.

[0178] In one embodiment, each prompt contains historical time-series data for n time points, where n is an integer greater than or equal to 1.

[0179] In one embodiment, the first processor 20 is further configured to: predict the time series data at the (n+1)th time point based on the historical time series data at n time points contained in each prompt; perform calculations on the predicted data and the historical time series data at the (n+1)th time point to obtain a first value; determine a second value based on the historical time series data at the nth time point and the historical time series data at the (n+1)th time point; and determine the score corresponding to each prompt in the prompt set based on the first value and the second value.

[0180] In one embodiment, the first processor 20 is further configured to recombine multiple highest-rated prompts to obtain M combined prompts, where M is an integer greater than or equal to 0; perform word order mutation processing on the M combined prompts to obtain M word order adjusted prompts; and construct an updated prompt set based on the multiple highest-rated prompts and the M word order adjusted prompts, the updated prompt set being used to select a target prompt from the updated prompt set.

[0181] In one embodiment, the first processor 20 is further configured to construct question-answer pairs based on historical text data; and to train the initial model based on the target prompt, historical time-series data, and question-answer pairs to obtain the target model.

[0182] Based on the above embodiments, another embodiment of this application provides a network fault prediction device 3, such as... Figure 12 As shown, the network fault prediction device 3 includes:

[0183] The second acquisition unit 30 is used to acquire user operation information and time sequence data associated with the network at the first moment.

[0184] The fusion unit 31 is used to fuse user operation information and time sequence data to obtain the input prompt.

[0185] The fault prediction unit 32 is used to predict network faults based on the input prompt and the target model, which is trained using the model training method described above.

[0186] In one embodiment, the network fault prediction device 3 may further include a conversion unit.

[0187] The conversion unit is used to convert and process user operation information to form the first input prompt.

[0188] The second acquisition unit 30 is also used to acquire a prompt template and, based on the prompt template, determine the second prompt to be input corresponding to the timing data.

[0189] The fusion unit 31 is also used to fuse the first input prompt and the second input prompt to obtain the input prompt.

[0190] In one embodiment, the fusion unit 31 is further configured to fuse and output the prediction results, user operation information, and alarm information obtained after network fault prediction using the target model.

[0191] This application provides a network fault prediction device that acquires user operation information and network-related time-series data at a first moment; fuses the user operation information and time-series data to obtain a prompt to be input; and performs network fault prediction based on the prompt and a target model, which is trained using the aforementioned model training method. Therefore, the network fault prediction device provided in this application, when using a target model trained based on historical time-series data for network fault prediction, can match the actual network situation at different times by adjusting the user operation information and time-series data input at different times, thus improving the accuracy of network fault prediction.

[0192] Figure 13This is a schematic diagram of the composition structure of a network fault prediction device 4 provided in an embodiment of this application. In practical applications, based on the same disclosed concept of the above embodiments, such as... Figure 13 As shown, the network fault prediction device 4 in this embodiment of the application includes a second processor 40, a second memory 41, and a second communication bus 42.

[0193] In specific embodiments, the second acquisition unit 30, fusion unit 31, fault prediction unit 32, and conversion unit described above can be implemented by a second processor 40 located on the network fault prediction device 4. The second processor 40 can be at least one of ASIC, DSP, DSPD, PLD, FPGA, CPU, controller, microcontroller, and microprocessor. It is understood that for different devices, the electronic device used to implement the above processor functions can also be other types, and this application embodiment does not impose specific limitations.

[0194] In this embodiment, the second communication bus 42 is used to realize the connection and communication between the second processor 40 and the second memory 41; when the second processor 40 executes the running program stored in the second memory 41, it implements the following network fault prediction method:

[0195] The system acquires user operation information and network-related time-series data at the first moment; it fuses the user operation information and time-series data to obtain the input prompt; and it performs network fault prediction based on the input prompt and the target model, which is trained using the above model training method.

[0196] In one embodiment, the second processor 40 is further configured to convert and process user operation information to form a first input prompt; obtain a prompt template, and determine a second input prompt corresponding to the timing data based on the prompt template; and fuse the first input prompt and the second input prompt to obtain the input prompt.

[0197] In one embodiment, the second processor 40 is further configured to fuse and output the prediction results, user operation information, and alarm information obtained after network fault prediction using the target model.

[0198] Based on the above embodiments, this application provides a storage medium storing a computer program thereon. The computer-readable storage medium stores one or more programs, which can be executed by one or more processors. The computer program can be applied to a model training device to implement the model training method described above, or applied to a network fault prediction device to implement the network fault prediction method described above.

[0199] Based on the above embodiments, this application provides a computer program product, including a computer program that can be executed by one or more processors and applied in a model training device to implement the model training method described above, or applied in a network fault prediction device to implement the network fault prediction method described above.

[0200] It should be noted that, in the embodiments of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0201] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to the related technology, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause an image display device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of the embodiments of this application.

[0202] The above description is merely a specific implementation of the embodiments of this application, but the protection scope of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A model training method, characterized in that, The method comprises: acquiring historical time series data and historical text data associated with a network; based on the historical time series data, constructing a prompt set; selecting a target prompt from the prompt set, the target prompt being superior to the remaining prompts in the prompt set other than the target prompt; based on the target prompt, the historical time series data and the historical text data, training an initial model to obtain a target model, the target model being used for predicting network failures; wherein the selecting a target prompt from the prompt set comprises: scoring each prompt in the prompt set; selecting multiple prompts from the prompt set multiple times each time; for each selected multiple prompts, selecting a prompt with the highest score from each selected multiple prompts to obtain multiple prompts with the highest score; selecting a target prompt from the multiple prompts with the highest score; the scoring each prompt in the prompt set comprises: based on the historical time series data of n time points contained in each prompt, predicting the time series data of the n+1 time point to obtain prediction data, n being an integer greater than or equal to 1; performing operation processing on the prediction data and the historical time series data of the n+1 time point to obtain a first value; and based on the historical time series data of the n time points and the historical time series data of the n+1 time point, determining a second value; based on the first value and the second value, determining the score corresponding to each prompt in the prompt set.

2. The method of claim 1, wherein, After the selecting a prompt with the highest score from each selected multiple prompts to obtain multiple prompts with the highest score, the method further comprises: recombining the multiple prompts with the highest score to obtain M combined prompts, M being an integer greater than or equal to 0; performing syntax sequence mutation processing on the M combined prompts to obtain M syntax sequence adjusted prompts; based on the multiple prompts with the highest score and the M syntax sequence adjusted prompts, constructing an updated prompt set, the updated prompt set being used to select a target prompt from the updated prompt set.

3. The method of claim 1, wherein, The training an initial model based on the target prompt, the historical time series data and the historical text data to obtain a target model comprises: based on the historical text data, constructing a question and answer pair; based on the target prompt, the historical time series data and the question and answer pair, training the initial model to obtain the target model.

4. A network failure prediction method characterized by, The method comprises: acquiring user operation information at a first time and time series data associated with a network at the first time; performing fusion processing on the user operation information and the time series data to obtain a to-be-input prompt; based on the to-be-input prompt and a target model, performing network failure prediction, the target model being trained by the method of any one of claims 1 to 3.

5. The method of claim 4, wherein, The fusion processing of the user operation information and the timing data obtains a to-be-input prompt, and the fusion processing comprises the following steps: The user operation information is converted to form a first to-be-input prompt; A prompt template is obtained, and a second to-be-input prompt corresponding to the timing data is determined based on the prompt template; The first to-be-input prompt and the second to-be-input prompt are fused to obtain a to-be-input prompt.

6. The method of claim 4, wherein, After the network fault prediction based on the to-be-input prompt and the target model, the method further comprises the following steps: The prediction result obtained after the network fault prediction by using the target model, the user operation information and the alarm information are fused and output.

7. A model training apparatus characterized by comprising: The model training device comprises: A first acquisition unit is configured to acquire historical timing data and historical text data associated with a network; A selection unit is configured to construct a prompt set based on the historical timing data, and select a target prompt from the prompt set, wherein the target prompt is better than the remaining prompts in the prompt set except the target prompt; A training unit is configured to train an initial model based on the target prompt, the historical timing data and the historical text data to obtain a target model, wherein the target model is used for predicting network faults; The scoring unit is configured to score each prompt in the prompt set; The selection unit is further configured to select multiple prompts from the prompt set for multiple times, select a prompt with the highest score from the multiple prompts selected each time, and obtain multiple prompts with the highest score, and select the target prompt from the multiple prompts with the highest score; A prediction unit is configured to predict timing data at an n+1th time point based on historical timing data at n time points included in each prompt, wherein n is an integer greater than or equal to 1, and obtain prediction data; A processing unit is configured to perform operation processing on the prediction data and historical timing data at the n+1th time point to obtain a first value, and determine a second value based on the historical timing data at the n time points and the historical timing data at the n+1th time point; The scoring unit is further configured to determine the score of each prompt in the prompt set based on the first value and the second value.

8. A network failure prediction apparatus characterized by comprising: The network fault prediction device comprises: A second acquisition unit is configured to acquire user operation information at a first time and timing data associated with the network at the first time; A fusion unit is configured to fuse the user operation information and the timing data to obtain a to-be-input prompt; A fault prediction unit is configured to predict network faults based on the to-be-input prompt and a target model, wherein the target model is trained by using the method of any one of claims 1 to 3.

9. A model training device comprising: The model training device comprises a first processor and a first memory, and the first processor implements the method of any one of claims 1 to 3 when executing a running program stored in the first memory.

10. A network failure prediction device characterized by comprising: The network failure prediction device comprises a second processor and a second memory; the second processor implements the method of any one of claims 4 to 6 when executing a running program stored in the second memory.

11. A storage medium having stored thereon a computer program, characterized in that The computer program, when executed by a processor, implements the method of any one of claims 1 to 3, or the computer program, when executed by a processor, implements the method of any one of claims 4 to 6.

12. A computer program product comprising a computer program, characterized in that, The computer program, when executed by a processor, implements the method of any one of claims 1 to 3, or the computer program, when executed by a processor, implements the method of any one of claims 4 to 6.

Images